Anton Blanchard anton at samba.org
Tue Jan 6 11:52:32 EST 2004

Hi Joel,

> Just got back from vacation so I'm not sure what the status of this is.
> Dec 16, before I was on vacation, I sent a patch to the list to try to fix
> locks as well, much of it is relavant to this discussion though it probably
> needs updating now as locks have been changing underneath it.
> I think we additionally need to seperate out HMT_LOW and HMT_MEDIUM.  It
> just happens that almost all of the current pSeries (and the G5 from
> Apple) don't support HMT, so it is a bit wasteful to call it.

Yeah I pushed the HMT_LOW/HMT_MEDIUM bits into the out of line slow path
based on your earlier results. My theory was the improvement you saw
came mainly from less icache footprint, so putting them out of line
in the slow and hopefully uncommon path should do the same.

Assuming we have to have a single kernel image for all pseries/g5
platforms, then we cant do a lot about them other than nop'ing them out.
Of course or r1,r1,r1 is a nop already although its not the preferred
nop (preferred nop does get handled a little more efficiently on POWER4
from memory).


** Sent via the linuxppc64-dev mail list. See http://lists.linuxppc.org/

More information about the Linuxppc64-dev mailing list